Distribution transformer weight and overload prediction method based on IAR-CI model

Abstract

The invention discloses a distribution transformer heavy overload prediction method based on an IAR-CI model, and the method comprises the steps: firstly, considering internal and external factors, collecting a plurality of data sources, building a distribution transformer operation state database, and recognizing a rare high-risk time period and a high-risk rare factor which are strongly correlated with the distribution transformer heavy overload through ICA, so as to cover extreme factors and scenes; secondly, designing a factor weight calculation method based on key importance measurement calculation so as to accurately measure the risk weight of the factor; and finally, applying a TBFP-Growth algorithm to enhance the operation efficiency of the model. According to the method, potential laws and relations in a large amount of offset data can be directly mined, the prediction result can be generated based on incomplete or short-span input data, and normalization and other preparatory work are not needed.

Description

Technical Field

[0001] The invention relates to a distribution transformer heavy overload prediction method based on an IAR-CI model. Background Art

[0002] As a key equipment in the power system, the transformer plays a vital role in power transmission and distribution. However, the transformer faces the problem of heavy overload during operation. When the transformer is overloaded, the internal heating of the transformer will increase, which may cause equipment damage, shorten equipment life or even fire, and cause power failures, voltage fluctuations and other problems. These potential problems not only directly affect the residents' satisfaction with electricity use, but also threaten people's production and life, and also bring great risks to the safe operation of the power grid. Therefore, heavy overload prediction is a key research field in the current power system.

[0003] With the development of digitalization and intelligence of power systems, distribution transformer heavy overload prediction has also become one of the key technologies for realizing intelligent condition maintenance. In recent years, machine learning methods have been widely used in heavy overload prediction. Some heavy overload prediction methods use machine learning algorithms to learn the relationship between transformer overload and other input factors and make predictions. However, when considering scenarios with multiple source factors, many machine learning methods do not make relevant analysis in the context of temporal and spatial data bias in real scenarios, resulting in some of the HRR factors being ignored, which will cause irreversible damage to the transformer once they appear. In addition, the adjustment and optimization of machine learning models may also be difficult. Therefore, how to solve the problem of some HRR factors being ignored in heavy overload prediction due to the lack of temporal and spatial data bias analysis in real scenarios is of great research value. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a distribution transformer heavy overload prediction method based on the IAR-CI model with a simple and reliable algorithm.

[0005] The technical solution of the present invention to solve the above technical problem is: a distribution transformer heavy overload prediction method based on the IAR-CI model, comprising the following steps:

[0006] Step 001: Establish a distribution transformer status feature library based on internal and external factors of distribution transformer heavy overload;

[0007] Step 002: Based on the training data of the database, the importance scores of association rules in each time period are calculated for all characteristic factors contained in each feature, and the high-frequency association sets and rare data sets in each feature are mined by comparing the corresponding calculated association state thresholds;

[0008] Step 003: Fault records containing any rare factor in the feature are classified into rare data sets, and HRR factors based on rare data sets are further mined;

[0009] Step 004: Calculate the relative risk weight of each factor through CI importance diagnosis;

[0010] Step 005: Run the simulation using the TBFP-Growth mining algorithm to enhance the model's operating efficiency.

[0011] In the above distribution transformer heavy overload prediction method based on the IAR-CI model, the specific process of step 001 is as follows:

[0012] When establishing the distribution transformer status feature library, it is necessary to integrate the internal and external factors of the distribution transformer overload and summarize them into a unified data processing space. The data processing space matrix construction process is as follows:

[0013] Assume that T = {t1, t2, …, t j ,…,t n} is a set containing all input features, where t j is the input feature, j = 1, 2, ... n, where each input feature t j are composed of a set of factors c j,1 ,c j,2 ,…,c j,k ,…,c j,l Composition, c j,k Represents the input feature t j The kth factor of , the total number of factors is l; let N = {n1,n2,…,n i ,…,n m} is the label corresponding to m records; let Y = {Y1, Y2, …, Y i ,…,Y m} is a set of m target variables, where Y i represents the i-th target variable, indicating whether the i-th transformer is overloaded, i = 1, 2, ... m; let P = {P1, P2, ..., P i ,…,P m} are m sets containing all the quarters where the faults are located, and then each set is written in matrix form: N = {n1, n2, ..., n i ,…,n m} T , T={t1,t2,…,t j ,…,t n}, Y={Y1,Y2,…,Y i ,…,Y m} T , P={P1,P2,…P i,…,P m} T Thus, a processing space matrix DB for distribution transformer heavy overload prediction is constructed. s ,as follows:

[0014]

[0015] Each row starting from the second row represents a record of a failure, t1,…,t j represents the characteristics of external factors, t j+1 ,…,t n represents the internal factor characteristics, m ij Representative record n i Medium feature t j An environmental factor C j,k ,Y i It is the result of multiple factors. i is the number label of this fault event, where P i represents the quarter in which the i-th failure occurs, that is, P i =P(s)∈{P(1),P(2),P(3),P(4)} represents any one of the four quarters.

[0016] In the above-mentioned distribution transformer overload prediction method based on the IAR-CI model, in step 001, the external factors include natural environmental factors and factors of the rest of the power system, wherein the natural environmental factors include ambient temperature, topography, and influence of flora and fauna, and the factors of the rest of the power system include load conditions, grid faults, and supply voltage quality, and the internal factors include transformer rated capacity, winding type, cooling method, and insulation material.

[0017] In the above-mentioned distribution transformer heavy overload prediction method based on the IAR-CI model, in step 002, support, confidence, Kulc metric and imbalance ratio IR are selected as association rule importance indicators, and the calculation formulas of each indicator are as follows:

[0018] For an association rule X→Y, X represents a subset of the input features, and the support formula of X is:

[0019]

[0020] B f () represents DB s The cardinality of fault records that simultaneously meet all the conditions included;

[0021] The confidence formula for an association rule X→Y is:

[0022]

[0023] The Kulc coefficient formula for an association rule X→Y is:

[0024]

[0025] The imbalance ratio formula of an association rule X→Y is:

[0026]

[0027] Support represents the processing space matrix DB s The proportion of records containing X to the total number in the input database DB s In , the number of item sets containing both X and Y accounts for the proportion of records containing X. The Kulc metric is used to measure the quality of the rule, and the imbalance ratio IR is used to measure the imbalance between the two item sets in the rule.

[0028] In the above-mentioned distribution transformer heavy overload prediction method based on the IAR-CI model, in step 002, in order to consider the importance of the attribute period, the associated state threshold is improved, specifically: the collected database data is divided into attribute periods according to time intervals, wherein a quarter in a year is defined as a benchmark attribute period; and the associated state threshold is set according to the importance of each attribute period, that is, based on the four association rule importance indicators, four corresponding attribute period importance thresholds are set for the distribution of faults in each season, and the corresponding threshold setting expression is:

[0029] Support threshold:

[0030]

[0031] DB s (i,n+3) means the i-th row and n+3-th column in the matrix, representing seasonal data;

[0032] Confidence Threshold:

[0033] minconf s =minconf0

[0034] Kulc metric threshold:

[0035]

[0036] Imbalance ratio threshold:

[0037]

[0038] Where i = 2, 3, ..., m represents the input database DB s One row, B f () represents DB sThe cardinality of fault records that simultaneously meet all the conditions included in the table, subscript 0 represents the initial set association state threshold, P(s) represents any season in spring, summer, autumn and winter, where: s∈{1,2,3,4}, P max (s) represents the quarter with the highest frequency of transformer overload, and Y(c) represents whether the transformer is overloaded or not; for all the characteristic factors contained in each feature, only the importance scores of the association rules in each time period reach or exceed the threshold of the corresponding time period and are considered in the high-frequency association set, while those below the threshold of the time period will be classified into the rare data set.

[0039] In the above-mentioned distribution transformer heavy overload prediction method based on the IAR-CI model, in step 003, the importance scores of association rules of rare factors in different forms are calculated according to the distribution of rare factors under different environmental characteristics, so that HRR factors strongly associated with the target can be further mined from the rare data set at the same time. For this purpose, the standard score calculation method of the four association indicators of support, confidence, Kulc coefficient and imbalance ratio IR is improved. When an association rule X c +X h →Y contains a certain environmental feature t j When any rare environmental element is j The calculation formula for the importance score of association rules with rare factors is:

[0040]

[0041]

[0042]

[0043]

[0044] Among them, S j(supp) , S j(conf) , S j(Kulc) , S j(IR) They represent the importance scores of the four correlation indicators, namely, the support, confidence, Kulc coefficient and imbalance ratio IR of the rare factor, respectively. c and X h Represent high-frequency association sets and rare data sets, respectively, M d Represents a numerical range from 2 to (n+1); if the association rule importance score of a rare factor reaches or exceeds the threshold, it is considered as an HRR factor, and if it is lower than the threshold, it is excluded.

[0045] In the above distribution transformer overload prediction method based on the IAR-CI model, in step 004, a weight calculation method for each factor under different environments is designed, and the specific calculation process is as follows:

[0046] A factor c j,k ∈t j Relative risk weight It consists of two parts, expressed as:

[0047]

[0048] in Represents the risk generated by high-frequency correlation factors; Indicates the risk generated by the HRR factor;

[0049] The risk calculation formula generated by high-frequency correlation factors is:

[0050]

[0051] Among them B f (m) stands for DB s The cardinality of all fault records in;

[0052] The critical importance CI measures the input factor c when the system determines that a fault has occurred. j,k The contribution to system failure is calculated as follows:

[0053]

[0054] In the formula, I cR (c j,k ) represents the risk of overall system failure, R i (n) represents the risk of component failure, and h(R(n)) represents the risk of overall system failure. By analyzing the CI of components, the potential impact of HRR factors on the system can be quantified.

[0055] I B The Birnbaum importance is the importance measure, and the calculation formula is:

[0056]

[0057] Where h(1 i , R(n)) represents when the factor c j,k When it is determined that there is a correlation, that is, the risk factor of the component is 1, the risk of the entire system failing; h(0 i ,p(n)) represents when the factor c j,k When it is determined that there is no correlation, that is, the risk of the component factor is 0, the risk of the entire system failing; t i represents a fault record, i.e., the i-th fault, and t j represents the jth environmental factor when the fault occurs;

[0058] The occurrence of a fault record is composed of unique elements under different characteristics. In the known fault records, if any characteristic condition is missing, the fault will not occur. Assuming that each characteristic factor is independent of each other, the risk of overall system failure is written as:

[0059]

[0060] Where X hh is a high-risk rare variable subspace containing variables with a certain feature t j The set of transactions with any HRR factor in ;

[0061] The risk of HRR factor is expressed as:

[0062]

[0063] In the above-mentioned distribution transformer heavy overload prediction method based on the IAR-CI model, in step 005, the applied TBFP-Growth algorithm adopts a heuristic strategy to select a conditional pattern base, and the specific steps of the mining algorithm are as follows:

[0064] (1) Input database, the items in each transaction are sorted in lexicographic order;

[0065] (2) Sort the factors in the input database according to their importance diagnostic criteria scores, traverse the path above the node on the current side link, and update the count on the path to generate a list of fuzzy frequent itemsets;

[0066] (3) Insert the scores of the importance diagnostic criteria into the FP-tree in descending order;

[0067] (4) Construct the conditional pattern library and subtree FP-tree;

[0068] (5) Based on the high-frequency factor set of the variable, the search space is pruned using both the minimum support constraint and the new constraint, and the frequent fuzzy sets of the corresponding part of the FP-tree are calculated;

[0069] (6) Recursively mine the corresponding high-frequency factor set;

[0070] (7) Determine whether the factors in the variable header have been processed. If they have been processed, generate a set of all high-frequency factors. If they have not been processed, repeat steps (5) to (7) until the end.

[0071] Beneficial effects of the present invention: The present invention first considers internal and external factors, collects multiple data sources and establishes a distribution transformer operation status database, and identifies rare high-risk periods and high-risk rare (HRR) factors that are strongly associated with heavy overload of distribution transformers through ICA to cover extreme factors and scenarios; secondly, based on the criticality importance (CI) metric calculation, a factor weight calculation method is designed to accurately measure the risk weight of the factor; finally, the TBFP-Growth algorithm is applied to enhance the operating efficiency of the model. The method of the present invention can directly mine the potential laws and connections in a large amount of biased data, and can generate prediction results based on incomplete or short-span input data without the need for preparatory work such as normalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of the present invention.

[0073] Figure 2 Schematic diagram of internal and external influencing factors of heavy overload of distribution transformer.

[0074] Figure 3 This is the flow chart of TBFP-Growth algorithm.

[0075] Figure 4 This is a comparison chart for algorithm testing.

[0076] Figure 5 This is a comparison chart of internal test results.

[0077] Figure 6 This is a comparison chart of external test results. DETAILED DESCRIPTION

[0078] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0079] like Figure 1 As shown, a distribution transformer heavy overload prediction method based on the IAR-CI model includes the following steps:

[0080] Step 001: Establish a distribution transformer status feature library based on internal and external factors of distribution transformer heavy overload.

[0081] In order to comprehensively predict the overload risk of distribution transformers, it is necessary to comprehensively consider the impact of internal and external factors on them. The combined effect of these internal and external factors determines the risk level of distribution transformer overload. Therefore, in order to ensure the comprehensiveness and accuracy of the model, the present invention comprehensively considers the impact of external and internal factors when constructing the distribution transformer state feature library, such as Figure 2As shown in the figure, external factors mainly include natural environmental factors and other factors of the power system. Natural environmental factors mainly include ambient temperature, topography, influence of flora and fauna, etc.; other factors of the electrical system mainly include load conditions, grid faults, power supply voltage quality, etc. Internal factors mainly include transformer rated capacity, winding type, cooling method, insulation material, etc.

[0082] The specific process of step 001 is:

[0083] When establishing the distribution transformer status feature library, it is necessary to integrate the internal and external factors of the distribution transformer overload and summarize them into a unified data processing space. The data processing space matrix construction process is as follows:

[0084] Assume that T = {t1, t2, …, t j ,…,t n} is a set containing all input features, where t j is the input feature, j = 1, 2, ... n, where each input feature t j are composed of a set of factors c j,1 ,c j,2 ,…,c j,k ,…,c j,l Composition, c j,k Represents the input feature t j The kth factor of , the total number of factors is l; let N = {n1,n2,…,n i ,…,n m} is the label corresponding to m records; let Y = {Y1, Y2, …, Y i ,…,Y m} is a set of m target variables, where Y i represents the i-th target variable, indicating whether the i-th transformer is overloaded, i = 1, 2, ... m; let P = {P1, P2, ..., P i ,…,P m} are m sets containing all the quarters where the faults are located, and then each set is written in matrix form: N = {n1, n2, ..., n i ,…,n m} T , T={t1,t2,…,t j ,…,t n}, Y={Y1,Y2,…,Y i ,…,Y m} T , P={P1,P2,…P i ,…,P m} T Thus, a processing space matrix DB for distribution transformer heavy overload prediction is constructed. s ,as follows:

[0085]

[0086] Each row starting from the second row represents a record of a failure, t1,…,t j represents the characteristics of external factors, t j+1 ,…,t n represents the internal factor characteristics, m ij Representative record n i Medium feature t j An environmental factor C j,k ,Y i It is the result of multiple factors. i is the number label of this fault event, where P i represents the quarter in which the i-th failure occurs, that is, P i =P(s)∈{P(1),P(2),P(3),P(4)} represents any one of the four quarters.

[0087] Step 002: Based on the training data in the database, the importance scores of association rules in each time period are calculated for all characteristic factors contained in each feature, and the corresponding calculated association state thresholds are compared to mine high-frequency association sets and rare data sets in each feature.

[0088] In step 002, support, confidence, Kulc metric and imbalance ratio IR are selected as association rule importance indicators, and the calculation formulas of each indicator are as follows:

[0089] For an association rule X→Y, X represents a subset of the input features, and the support formula of X is:

[0090]

[0091] B f () represents DB s The cardinality of fault records that simultaneously meet all the conditions included;

[0092] The confidence formula for an association rule X→Y is:

[0093]

[0094] The Kulc coefficient formula for an association rule X→Y is:

[0095]

[0096] The imbalance ratio formula of an association rule X→Y is:

[0097]

[0098] Support represents the processing space matrix DB s The proportion of records containing X to the total number in the input database DB s In , the number of item sets containing both X and Y accounts for the proportion of records containing X. The Kulc metric is used to measure the quality of the rule, and the imbalance ratio IR is used to measure the imbalance between the two item sets in the rule.

[0099] In order to consider the importance of the attribute period, the association state threshold is improved. Specifically, the collected database data is divided into attribute periods according to time intervals, where one quarter of a year is defined as the benchmark attribute period; and the association state threshold is set according to the importance of each attribute period, that is, based on the four association rule importance indicators, four corresponding attribute period importance thresholds are set for the distribution of faults in each season. The corresponding threshold setting expression is:

[0100] Support threshold:

[0101]

[0102] DB s (i,n+3) means the i-th row and n+3-th column in the matrix, representing seasonal data;

[0103] Confidence Threshold:

[0104] minconf s =minconf0

[0105] Kulc metric threshold:

[0106]

[0107] Imbalance ratio threshold:

[0108]

[0109] Where i = 2, 3, ..., m represents the input database DB s One row, B f () represents DB s The cardinality of fault records that simultaneously meet all the conditions included in the table, subscript 0 represents the initial set association state threshold, P(s) represents any season in spring, summer, autumn and winter, where: s∈{1,2,3,4}, P max(s) represents the quarter with the highest frequency of transformer overload, and Y(c) represents whether the transformer is overloaded or not. For all the characteristic factors contained in each feature, only the importance scores of the association rules in each time period reach or exceed the threshold of the corresponding time period and are considered in the high-frequency association set, while those below the threshold of the time period will be classified into the rare data set. By introducing the threshold of the importance of the attribute time period to improve the calculation, the model can more effectively discover the laws and trends related to transformer overload failures.

[0110] Step 003: Fault records containing any rare factors in the feature are classified into rare data sets, and HRR factors based on rare data sets are further mined.

[0111] In step 003, according to the distribution of rare factors under different environmental characteristics, the importance scores of association rules of rare factors in different forms are calculated, so that HRR factors strongly associated with the target can be further mined from the rare data set at the same time. For this purpose, the standard score calculation method of the four association indicators of support, confidence, Kulc coefficient and imbalance ratio IR is improved. When an association rule X c +X h →Y contains a certain environmental feature t j When any rare environmental element is j The calculation formula for the importance score of association rules with rare factors is:

[0112]

[0113]

[0114]

[0115]

[0116] Among them, S j(supp) , S j(conf) , S j(Kulc) , S j(IR) They represent the importance scores of the four correlation indicators, namely, the support, confidence, Kulc coefficient and imbalance ratio IR of the rare factor, respectively. c and X h Represent high-frequency association sets and rare data sets, respectively, M d Represents a numerical range from 2 to (n+1); if the association rule importance score of a rare factor reaches or exceeds the threshold, it is considered as an HRR factor, and if it is lower than the threshold, it is excluded.

[0117] Step 004: Calculate the relative risk weight of each factor through CI importance diagnosis.

[0118] In step 004, the relative risk weights of each element are determined according to their influence on the system. In the component importance measurement, CI is an indicator used to measure the importance of a component to the overall function and performance of the system. Based on the critical importance CI, a weight calculation method for each factor in different environments is designed. The specific calculation process is as follows:

[0119] A factor c j,k ∈t j Relative risk weight It consists of two parts, expressed as:

[0120]

[0121] in Represents the risk generated by high-frequency correlation factors; Indicates the risk generated by the HRR factor;

[0122] The risk calculation formula generated by high-frequency correlation factors is:

[0123]

[0124] Among them B f (m) stands for DB s The cardinality of all fault records in;

[0125] The critical importance CI measures the input factor c when the system determines that a fault has occurred. j,k The contribution to system failure is calculated as follows:

[0126]

[0127] In the formula, I CR (c j,k ) represents the risk of overall system failure, R i (n) represents the risk of component failure, and h(R(n)) represents the risk of overall system failure. By analyzing the CI of components, the potential impact of HRR factors on the system can be quantified.

[0128] I B The Birnbaum importance is the importance measure, and the calculation formula is:

[0129]

[0130] Where h(1 i , R(n)) represents when the factor c j,k When it is determined that there is a correlation, that is, the risk factor of the component is 1, the risk of the entire system failing; h(0 i ,p(n)) represents when the factor c j,kWhen it is determined that there is no correlation, that is, the risk of the component factor is 0, the risk of the entire system failing; t i represents a fault record, i.e., the i-th fault, and t j represents the jth environmental factor when the fault occurs;

[0131] The occurrence of a fault record is composed of unique elements under different characteristics. In the known fault records, if any characteristic condition is missing, the fault will not occur. Assuming that each characteristic factor is independent of each other, the risk of overall system failure is written as:

[0132]

[0133] Where X hh is a high-risk rare variable subspace containing variables with a certain feature t j The set of transactions with any HRR factor in ;

[0134] The risk of HRR factor is expressed as:

[0135]

[0136] Step 005: Run the simulation using the TBFP-Growth mining algorithm to enhance the model's operating efficiency.

[0137] In step 005, the applied TBFP-Growth algorithm adopts a heuristic strategy to select the conditional pattern base. The TBFP-Growth algorithm adopts a heuristic strategy to select the conditional pattern base, which avoids scanning all frequent items one by one, thereby reducing the number of recursive calls. This improvement means that the algorithm no longer needs to scan the entire data set for frequent items, but only needs to build a conditional FP tree. This optimization not only improves the efficiency of the algorithm, but also reduces memory consumption, especially for mining tasks in large-scale data sets and resource-constrained environments, it shows better scalability. Compared with the FP-Growth algorithm, the TBFP-Growth algorithm has made significant progress in improving algorithm performance and efficiency. By reducing the number of times the conditional pattern base is recursively constructed, the algorithm can effectively speed up the mining process. The specific process of the TBFP-Growth algorithm is as follows Figure 3 As shown, the steps are as follows:

[0138] (1) Input database, the items in each transaction are sorted in lexicographic order;

[0139] (2) Sort the factors in the input database according to their importance diagnostic criteria scores, traverse the path above the node on the current side link, and update the count on the path to generate a list of fuzzy frequent itemsets;

[0140] (3) Insert the scores of the importance diagnostic criteria into the FP-tree in descending order;

[0141] (4) Construct the conditional pattern library and subtree FP-tree;

[0142] (5) Based on the high-frequency factor set of the variable, the search space is pruned using both the minimum support constraint and the new constraint, and the frequent fuzzy sets of the corresponding part of the FP-tree are calculated;

[0143] (6) Recursively mine the corresponding high-frequency factor set;

[0144] (7) Determine whether the factors in the variable header have been processed. If they have been processed, generate a set of all high-frequency factors. If they have not been processed, repeat steps (5) to (7) until the end.

[0145] The IAR-CI classifier model is uniformly used, and different algorithms are used for comparison. The comparison results are as follows Figure 4 shown.

[0146] Through comparative analysis Figure 5 It can be seen from the figure that the IAR-CI classifier model has a larger AUC value on the ROC and PR curves compared with the other two models. Therefore, the model based on the improved association rules and CI importance risk weight calculation method can achieve more accurate distribution transformer overload diagnosis.

[0147] Finally, the IAR-CI classifier model is compared with the currently common classifier models SVM, ELM, and NB to evaluate the performance of the distribution transformer heavy overload prediction model proposed in the present invention.

[0148] The ten-fold cross validation method was used to plot the ROC and PR curves of each classifier and calculate the AUC value of each model. Figure 6 ,Through simulation, it can be seen that the IAR-CI model can obtain the highest ,fault diagnosis accuracy, where the AUC values ​​of ROC and PR are 0.931 and 0.942 respectively, while the judgment ,accuracy of other classifier models is lower, and the AUC values ​​under the PR curve

Claims

1. A distribution transformer heavy overload prediction method based on IAR-CI model, characterized in that: The following steps are involved: Step 001: Establish a distribution transformer status feature library based on internal and external factors of distribution transformer heavy overload; Step 002: Based on the training data of the database, the importance scores of association rules in each time period are calculated for all characteristic factors contained in each feature, and the high-frequency association sets and rare data sets in each feature are mined by comparing the corresponding calculated association state thresholds; Step 003: Fault records containing any rare factor in the feature are classified into rare data sets, and HRR factors based on rare data sets are further mined; Step 004: Calculate the relative risk weight of each factor through CI importance diagnosis; Step 005: Run the simulation using the TBFP-Growth mining algorithm to enhance the model's operating efficiency.

2. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 1 is characterized in that: The specific process of step 001 is: When establishing the distribution transformer status feature library, it is necessary to integrate the internal and external factors of the distribution transformer overload and summarize them into a unified data processing space. The data processing space matrix construction process is as follows: Assume T = {t1, t2, ..., t j , ..., t n } is a set containing all input features, where t j is the input feature, j = 1, 2, ... n, where each input feature t j are composed of a set of factors c j，1 , c j，2 , ..., c j，k , ..., c j，l Composition, c j，k Represents the input feature t j The kth factor of , the total number of factors is l; let N = {n1, n2, ..., n i , ..., n m } is the label corresponding to m records; let Y = {Y1, Y2, ..., Y i , ..., Y m } is a set of m target variables, where Y i represents the i-th target variable, indicating whether the i-th transformer is heavily overloaded, i = 1, 2, ... m; let P = {P1, P2, ..., P i , ..., P m } are m sets containing all the quarters where the faults are located, and then each set is written in matrix form: N = {n1, n2, ..., n i , ..., n m } T , T={t1,t2,...,t j , ..., t n }, Y={Y1, Y2, ..., Y i , ..., Y m } T , P = {P1, P2, ... P i , ..., P m } T Thus, a processing space matrix DB for distribution transformer heavy overload prediction is constructed. s ,as follows: Each row starting from the second row represents a record of a failure, t1,...,t j represents the characteristics of external factors, t j+1 , ..., t n represents the internal factor characteristics, m ij Representative record n i Medium feature t j An environmental factor c j，k , Y i It is the result of multiple factors. i is the number label of this fault event, where P i represents the quarter in which the i-th failure occurs, that is, P i =P(s)∈{P(1), P(2), P(3), P(4)} represents any one of the four quarters.

3. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 2 is characterized in that: In step 001, the external factors include natural environmental factors and other factors of the power system, wherein the natural environmental factors include ambient temperature, topography, and influence of flora and fauna, the other factors of the power system include load conditions, grid faults, and supply voltage quality, and the internal factors include transformer rated capacity, winding type, cooling method, and insulation material.

4. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 2 is characterized in that: In step 002, support, confidence, Kulc metric and imbalance ratio IR are selected as association rule importance indicators, and the calculation formulas of each indicator are as follows: For an association rule X→Y, X represents a subset of the input features, and the support formula of X is: B f () represents DB s The cardinality of fault records that simultaneously meet all the conditions included; The confidence formula for an association rule X→Y is: The Kulc coefficient formula for an association rule X→Y is: The imbalance ratio formula of an association rule X→Y is: Support represents the processing space matrix DB s The proportion of records containing X to the total number in the input database DB s In , the number of item sets containing both X and Y accounts for the proportion of records containing X. The Kulc metric is used to measure the quality of the rule, and the imbalance ratio IR is used to measure the imbalance between the two item sets in the rule.

5. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 4 is characterized in that: In step 002, in order to consider the importance of the attribute period, the association state threshold is improved, specifically: the collected database data is divided into attribute periods according to time intervals, wherein one quarter of a year is defined as the benchmark attribute period; and the association state threshold is set according to the importance of each attribute period, that is, based on the four association rule importance indicators, four corresponding attribute period importance thresholds are set for the distribution of faults in each season, and the corresponding threshold setting expression is: Support threshold: DB s (i, n+3) means the i-th row and n+3-th column in the matrix, representing seasonal data; Confidence Threshold: minconf s =minconf0 Kulc metric threshold: Imbalance ratio threshold: Where i = 2, 3, ..., m represents the input database DB s One row, B f () represents DB s The cardinality of fault records that simultaneously meet all the conditions included in the , subscript 0 represents the initial set association state threshold, P(s) represents any season in spring, summer, autumn and winter, where: s∈{1, 2, 3, 4}, P max (s) represents the quarter with the highest frequency of transformer overload, and Y(c) represents whether the transformer is overloaded or not; for all the characteristic factors contained in each feature, only the importance scores of the association rules in each time period reach or exceed the threshold of the corresponding time period and are considered in the high-frequency association set, while those below the threshold of the time period will be classified into the rare data set.

6. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 5 is characterized in that: In step 003, according to the distribution of rare factors under different environmental characteristics, the importance scores of association rules of rare factors in different forms are calculated, so that HRR factors strongly associated with the target can be further mined from the rare data set at the same time. For this purpose, the standard score calculation method of the four association indicators of support, confidence, Kulc coefficient and imbalance ratio IR is improved. When an association rule X c +X h →Y contains a certain environmental feature t j When any rare environmental element is j The calculation formula for the importance score of association rules with rare factors is: Among them, S j(supp) , S j(conf) , S j(Kulc) , S j(IR) They represent the importance scores of the four correlation indicators, namely, the support, confidence, Kulc coefficient and imbalance ratio IR of the rare factor, respectively. c and X h Represent high-frequency association sets and rare data sets, respectively, M d Represents a numerical range from 2 to (n+1); if the association rule importance score of a rare factor reaches or exceeds the threshold, it is considered as an HRR factor, and if it is lower than the threshold, it is excluded.

7. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 6 is characterized in that: In step 004, a weight calculation method for each factor in different environments is designed, and the specific calculation process is as follows: A factor c j，k ∈t j Relative risk weight It consists of two parts, expressed as: in Represents the risk generated by high-frequency correlation factors; Indicates the risk generated by the HRR factor; The risk calculation formula generated by high-frequency correlation factors is: Among them B f (m) stands for DB s The cardinality of all fault records in; The critical importance CI measures the input factor c when the system determines that a fault has occurred. j，k The contribution to system failure is calculated as follows: In the formula, I CR (c j，k ) represents the risk of overall system failure, R i (n) represents the risk of component failure, and h(R(n)) represents the risk of overall system failure. By analyzing the CI of components, the potential impact of HRR factors on the system can be quantified. I B Birnbaum importance is the importance measure, and the calculation formula is: Where h(1 i , R(n)) represents when the factor c j，k When it is determined that there is a correlation, that is, the risk factor of the component is 1, the risk of the entire system failing; h(0 i , p(n)) represents when the factor c j，k When it is determined that there is no correlation, that is, the risk of the component factor is 0, the risk of the entire system failing; t i represents a fault record, i.e., the i-th fault, and t j represents the jth environmental factor when the fault occurs; The occurrence of a fault record is composed of unique elements under different characteristics. In the known fault records, if any characteristic condition is missing, the fault will not occur. Assuming that each characteristic factor is independent of each other, the risk of overall system failure is written as: Where X hh is a high-risk rare variable subspace containing variables with a certain feature t j The set of transactions with any HRR factor in ; The risk of HRR factor is expressed as:

8. The distribution transformer heavy overload prediction method based on the IAR-CI model according to claim 1 is characterized in that: In step 005, the applied TBFP-Growth algorithm adopts a heuristic strategy to select a conditional pattern base, and the specific steps of the mining algorithm are as follows: (1) Input database, the items in each transaction are sorted in lexicographic order; (2) Sort the factors in the input database according to their importance diagnostic criteria scores, traverse the path above the node on the current side link, and update the count on the path to generate a list of fuzzy frequent itemsets; (3) Insert the scores of the importance diagnostic criteria into the FP-tree in descending order; (4) Construct the conditional pattern library and subtree FP-tree; (5) Based on the high-frequency factor set of the variable, the search space is pruned using both the minimum support constraint and the new constraint, and the frequent fuzzy sets of the corresponding part of the FP-tree are calculated; (6) Recursively mine the corresponding high-frequency factor set; (7) Determine whether the factors in the variable header have been processed. If they have been processed, generate a set of all high-frequency factors. If they have not been processed, repeat steps (5) to (7) until the end.

Images